Skip to content

TroubleshootingComfyUITroubleshootingMiniMax H3Errors

MiniMax H3 ComfyUI error? Six messages, six fixes (2026)

· updated · 9 min read · 10 sources

Download counts as of , from each site

In short

Six MiniMax H3 messages cover most of what creators posted to ComfyUI's issue tracker: time_shift_slope is an outdated Spectrum node, aten::_int_mm on a Mac is fixed from v0.35.0, a black video on Apple Silicon has only workarounds, HostBuffer.read_file_slice failed is still open, NestedTensor in tiled decode was fixed in v0.32.0, and Sage Attention noise appears on RTX 50-series cards above roughly 160k tokens. Update first, then match your message.

A MiniMax H3 workflow that stops at the sampler is rarely a broken model file. In the ComfyUI issue tracker the same few messages keep coming back, and three of them have a documented fix that is only a version number away. The other three are open, so the honest answer for them is a workaround and a date.

This guide matches each message to its cause and fix, in the order that costs least time. Versions and fixes are read from Comfy's changelog and the linked pull request. What people reported comes from the issue threads themselves, read on 7 November 2026, and where a thread never reached an answer, the text says so. We did not reproduce any of these on our own hardware.

Update ComfyUI first, then read the message

The changelog shows how fast H3 support moved. v0.31.0 (7 August) fixed MiniMax H3 audio corruption with EasyCache and latent noise mask sampling, v0.32.0 (11 August) fixed a peak memory issue and a tiled decode crash, v0.35.0 (9 September) added the Mac int8 fix, and v0.40.0 (13 October) changed how single images are decoded. A build from before those dates reproduces errors that were already closed. The latest entry we read is v0.44.0 from 4 November.

docs.comfy.orgOpen the page
Comfy's changelog entry for v0.35.0, 9 September 2026, with the line MPS int8 highlighted: disable int8 weight-only quantization when torch._int_mm is missing
Comfy's own changelog: the Mac crash in the table below is closed from v0.35.0 onward. Source: ComfyUI changelog.

Match your message to the fix

What you seeCauseFix
RuntimeError: native MiniMax H3 module is missing required helpers: time_shift_slopeThe ComfyUI-Spectrum-MiniMax-H3 custom node expects a helper ComfyUI removed when it moved audio schedule handling into ModelSamplingAVUpdate that node to v0.1.6 or newer, restart ComfyUI (issue 15378, closed)
The operator 'aten::_int_mm' is not currently implemented for the MPS deviceThe int8 path calls torch._int_mm, which PyTorch lacks on MPSUpdate to v0.35.0 or newer; before that, set PYTORCH_ENABLE_MPS_FALLBACK=1 (issue 16127, closed)
Black video and Input contains (near) NaN/+-Inf on save, Apple Silicon or AMDNot established; reporters point at fp16 and at dynamic VRAM combinationsRemove --force-fp16, use the bf16 model, try --disable-dynamic-vram (issue 15315, no merged fix)
HostBuffer.read_file_slice failed, then CUDA error: out of memoryUnconfirmed; a maintainer suspects a CUDA bug that shows mostly with two GPUs--cuda-device 0, --disable-pinned-memory, or --disable-dynamic-vram (issue 15255, open)
to() received an invalid combination of arguments - got (NestedTensor) in tiled VAE decodeTiled decode passed a NestedTensor into a .to() callUpdate to v0.32.0 or newer; try the plain VAE Decode first (issue 15274)
Noise in video and audio with --use-sage-attention on an RTX 50-series cardSageAttention FP8 kernels degrade past a token count on sm_120; cause unconfirmedStay under about 151k tokens, or drop the flag for H3 (issue 15263, open)
Message, cause and fix, from the issue threads and Comfy's changelog (read 7 November 2026)

The six messages, one by one

time_shift_slope: it is the custom node, not ComfyUI

Opened on 7 August and closed the same day, issue 15378 reports a failure at SamplerCustomAdvanced. A commenter traced it to the Spectrum node: ComfyUI had moved audio schedule and velocity handling into ModelSamplingAV and removed the native helper that older Spectrum releases expected. The node's maintainer pointed to v0.1.6, which adds backward compatibility. The node's release page shows it has moved on to v0.2.29 (8 October), so update the node in ComfyUI Manager and restart. No reporter confirmed the fix in the thread, and a later comment says the same message appeared with the EasyCache node, which the v0.31.0 changelog line about EasyCache audio addresses.

aten::_int_mm on a Mac: fixed from v0.35.0

Issue 16127 came from ComfyUI Desktop 0.34.5 on an Apple machine with 48 GB of unified memory, failing at step 0 of 20. The cause, per pull request 16130: H3 checkpoints use int8 tensorwise weights, and that format reaches torch._int_mm, which MPS does not implement. The pull request makes ComfyUI check for support and fall back to the full precision path for those layers, so the model runs instead of crashing. That costs speed, and it was merged on 7 September and listed in v0.35.0.

Black video and NaN audio: only workarounds

Issue 15315 describes a completely black video from the official text to video workflow on an M4 Max, with a NaN error when audio was saved. Image to video and reference workflows worked on the same machine. The author closed it after porting attention to MLX and adding a custom audio guard node, which is a private fix, not a ComfyUI change. In the replies, one user fixed it by removing --force-fp16 and switching to the bf16 model, and others found --disable-dynamic-vram stable but slower. Reports of the same save error on AMD and Linux followed. Treat all of this as unproven: nothing in the thread shows a cause.

HostBuffer.read_file_slice failed: open, with a pattern

Issue 15255 opened on 3 August and has more than 70 comments, with activity as late as 16 October. It is assigned to a maintainer, who said the working theory is a CUDA bug. The reports cluster on machines where two GPUs are visible, though single-GPU cards appear too. Workarounds people reported: --cuda-device 0, --disable-pinned-memory, --disable-dynamic-vram (which one user then followed with an out of memory error in the int8 linear layer), and rolling back to 0.29.2 or 0.36. Our Dynamic VRAM guide covers the flags in more detail. There is no merged fix as of the last comment.

NestedTensor in tiled decode: update, or skip tiling

Issue 15274 shows a TypeError in the MiniMax VAE decode when the tiled node receives a NestedTensor. The changelog lists the repair under v0.32.0 as a fix for a crash on NestedTensor latents in VAEDecodeTiled. The issue itself is still open and labeled User Support, and the maintainer's advice there is to try the standard decoder first because the MiniMax VAE has built-in tiling and uses under 8 GB of VRAM.

Sage Attention noise on RTX 50-series

Issue 15263 reports noise in both video and audio with --use-sage-attention, with no warning printed. It was seen on an RTX PRO 6000 Blackwell and an RTX 5090, and the reporter's workaround is to stay below about 151k tokens (1920x832 at 15 seconds is their example) or to disable Sage. The cause is unconfirmed, and the reporter says a later version rendered a case that had failed, without claiming it as a fix.

The fix order that works

  1. 1

    Read your version in the startup log

    Compare it with the dates above. Anything before 0.35.0 can still hit the Mac crash, and anything before 0.32.0 the tiled decode crash.

  2. 2

    Update custom nodes, especially Spectrum

    A custom node pinned to an older ComfyUI is the cause of time_shift_slope. Update it, restart, then reload the page.

  3. 3

    Launch with the defaults

    Remove Sage, fp16 and pinned-memory flags. If the error goes away, add them back one at a time.

  4. 4

    Use the plain VAE Decode

    The MiniMax VAE tiles itself. Use the tiled node only if the plain decode really runs out of memory.

  5. 5

    Choose a different quant when memory fails

    Two threads report that the fp8 or int8 files avoided an out of memory error that the larger one hit. That is a report, not a rule.

  6. 6

    Still failing: search the issue by the exact message

    Open issues 15255 and 15263 are the live threads. Add your versions, hardware and launch flags there, not a screenshot of the error alone.

Triage path for a MiniMax H3 failure
Error at sampler or…ComfyUI 0.35.0 or n…Update ComfyUI and …Launch without opti…Plain VAE decodeReport on the open …

The order that clears the closed issues first and leaves the open ones for last.

Musetta studioDemo data
The Musetta studio home screen with a request typed for the character Mira, a 9:16 MiniMax H3 reel at golden hour
Staged with demo data: a reel request for Mira typed into the studio once the update is done. Nothing runs until you confirm.

What creators are saying

What creators are saying

  • A user on a Mac reported the MPS int8 crash at step 0 of 20 with custom nodes disabled, and the issue was closed two days later with a fix.

    ComplaintComfyUI issue 16127

  • Someone on an M4 Max got a black video from the text to video workflow while image to video worked, and later reports found the same save error on AMD.

    ComplaintComfyUI issue 15315

  • One user said removing the force-fp16 flag and switching to the bf16 model cured the black output, though generation stayed slow.

    TipComfyUI issue 15315

  • Several users on different RTX cards said the dynamic VRAM crash stopped when they used a single GPU or disabled pinned memory, while others still hit it.

    TipComfyUI issue 15255

  • A user with a 4 GB card hit the tiled decode crash and asked about a lighter decoder, which no one in the thread had tested.

    QuestionComfyUI issue 15274

What this does not fix

None of these fixes gives a model your character. H3 needs a reference or a LoRA for a recurring face, which the MiniMax H3 review and the Turbo workflow breakdown cover. If the message you have is a missing node, not an H3 failure, the missing nodes guide is the general routine, and if your card is simply too small, the rent-a-GPU page lists the alternatives. For MOSS Image errors in particular, see the MOSS-Image error guide. To train that LoRA on a 12 GB card, follow the ai-toolkit run order.

MiniMax H3 error cheat sheet

Mac int8 crash
Fixed in v0.35.0 (changelog, 9 Sep 2026)
Tiled decode NestedTensor
Fixed in v0.32.0 (changelog, 11 Aug 2026)
EasyCache audio corruption
Fixed in v0.31.0 (changelog, 7 Aug 2026)
time_shift_slope
Update ComfyUI-Spectrum-MiniMax-H3 to v0.1.6 or newer
HostBuffer.read_file_slice
Open: --cuda-device 0 or --disable-pinned-memory
Sage noise
Open: stay under about 151k tokens or drop the flag
Apple black video
Open: no --force-fp16, bf16 model

Verdict

Update before you debug. Two of the six messages are closed by a version, one by a custom node update, and the other three are open threads where the only evidence is what other creators tried. Do not treat a workaround from a thread as a fix: several are reported by a single user, and the dynamic VRAM crash has mixed results across versions.

Musetta runs on your own ComfyUI and installs the nodes and models a workflow is missing, so a mismatch can show up before a run rather than in the middle of one. It does not change which ComfyUI build you run, and it cannot fix a bug that is still open upstream. See how it compares with working in ComfyUI directly.

Questions

What does time_shift_slope mean in MiniMax H3?
It names a helper that older versions of the ComfyUI-Spectrum-MiniMax-H3 node expect and ComfyUI removed. In issue 15378 the fix is updating that node to v0.1.6 or newer and restarting.
Why does MiniMax H3 crash on my Mac with aten::_int_mm?
The int8 weights call torch._int_mm, which MPS does not implement. ComfyUI v0.35.0 added a fallback to full precision for those layers, per the changelog and pull request 16130.
Why is my MiniMax H3 video black on Apple Silicon?
The thread has no confirmed cause. One user fixed it by removing --force-fp16 and using the bf16 model, and others found --disable-dynamic-vram stable but slower.
Is the HostBuffer.read_file_slice failed error fixed?
Not in the thread we read. Issue 15255 was still open after its last comment on 16 October 2026, and users reported single-GPU, no pinned memory or no dynamic VRAM as workarounds.
Should I use the tiled VAE decode for MiniMax H3?
Try the plain VAE Decode first. In issue 15274 a maintainer says the MiniMax VAE has built-in tiling and uses under 8 GB of VRAM.

Sources

  1. 01ComfyUI changelog · Comfy docs, 2026-11-04
  2. 02Issue 15378: Minimax H3 SamplerCustomAdvance Error · GitHub, Comfy-Org, 2026-08-07
  3. 03Issue 16127: Comfy desktop issue with Minimax H3 · GitHub, Comfy-Org, 2026-09-05
  4. 04Pull request 16130: disable int8 without torch._int_mm · GitHub, Comfy-Org, 2026-09-07
  5. 05Issue 15315: black video and NaN audio on Apple Silicon · GitHub, Comfy-Org, 2026-08-05
  6. 06Issue 15255: HostBuffer.read_file_slice failed · GitHub, Comfy-Org, 2026-08-03
  7. 07Issue 15274: minimax h3 vae decode tiled error · GitHub, Comfy-Org, 2026-08-04
  8. 08Issue 15263: SageAttention noise on sm_120 · GitHub, Comfy-Org, 2026-08-03
  9. 09ComfyUI-Spectrum-MiniMax-H3 releases · GitHub, xmarre, 2026-10-08
  10. 10MiniMax H3 out of memory in ComfyUI: 16 GB fixes · FreakyJolly, 2026-11-06

Researched and written by the Musetta editorial desk with AI assistance, from the sources above. Numbers come from those sources; nothing is benchmarked by us unless the article says so. Spotted an error? Tell us on Discord and we fix it.